Edit by instruction

Edit by instruction only needs one sentence describing the change (new clothes, a new background, another season or style). The model redraws the image using it as a condition and keeps what you don’t mention as close to the original as possible. The result is saved as a new version of the image, which you can compare with the slider.

Required models

You need one of the following (put it in your model folder and set up its VAE and text encoder; see the Model guide):

Model	Notes	Default steps	Default CFG
Qwen Image 2.1	Can bring in other reference images	40	1
FLUX.2 klein	The fastest; can bring in other reference images	4	1
FLUX.1 Kontext	Accepts one condition image only; the file name must contain kontext to appear in the list	28	2.5

Choose the model under “Model”. Steps, CFG and sampler are remembered separately for each kind of model and reused next time.

Basic steps

1. Choose “Edit by instruction” as the refine mode.
2. In “What to change”, describe the change in one sentence, for example:
   • Change the background to a night city street
   • Make her wear a red hat
   • Turn it into a watercolor painting
3. Press “Edit”.

English is usually the most accurate; Chinese is understood too. English is recommended for FLUX.1 Kontext, for example Make it night.

Settings

Item	Description
Size	“Original”, or “About 1 MP”. The latter is faster, and the difference grows with the size of the original
Steps	More steps give more detail; fewer are faster
CFG	1 is fastest for Qwen and klein; higher follows the instruction more closely but doubles the time. For Kontext, CFG is the guidance value: higher follows the instruction more closely at no extra time
Sampler	Usually doesn’t need changing

Adding reference images

To bring in something from another image (clothes, a background, another person), press “+ Add” under “Other reference images”; up to 2 can be added.

• The image being edited is “Image 1”; the added ones are “Image 2” and “Image 3”.
• Refer to them by these names in the prompt (image 1, image 2 in English), for example “Put the dress from image 2 on the girl in image 1”.
• Each extra image makes it a little slower and uses a little more graphics memory.

FLUX.1 Kontext and reference images

Kontext accepts only one condition image. Extra references are joined to the right of the image being edited, so describe them by position instead (for example the dress on the right). The effect is weaker; to bring in content from another image, Qwen Image 2.1 or FLUX.2 klein is recommended.

Speed

Measured on an RTX 5080:

Model	Conditions	Time
FLUX.2 klein 4B	1024×1024, 4 steps	about 5 s
Qwen Image 2.1	1024×1344, 40 steps	about 29 s
FLUX.1 Kontext	768×768, 28 steps	about 52 s